Anytime Lab

Anytime World is a budget-slider demo of anytime computing. One scene is written once as a program — a canonical SDF scene in GFL, a budget-aware scene language — and rendered under a compute budget the user controls. Drag the budget slider from 1 to 100 and the world refines from a coarse estimate to a faithful reconstruction, while a live graph plots reconstruction error against budget, so every unit of compute shows exactly what it bought. Stop the slider anywhere and the image is complete and valid — that is the anytime property. The demo is an interactive four-region app (Target / Render / Controls / Metrics) that runs entirely locally: a WebGL2 raymarcher compiles the scene program, structural level-of-detail follows the budget, and the error curve is measured from real renders. The full architecture pairs this canonical scene with a neural side — Muse Image for target appearance, SAM 3.1 for the coarse first pass, and a Muse Spark-guided planner that would spend each budget unit on the highest-error region first. The submitted build is cache-first by design: preset targets and refinement ladders are precomputed, so the demo completes offline; those three models are its documented integration points rather than live calls. In the demo video, the uniform-allocation error curve is measured from the app's own renders; the guided curve shown next to it is a clearly labelled model-based illustration of error-first allocation, not a live planner run.